Vamos para um pouco de código o/¶
Imports¶
In [1]:
%matplotlib inline
import collections
import cv2
import dlib
from imutils import face_utils
from matplotlib import pyplot as plt
import numpy as np
# from google.colab.patches import cv2_imshow
Códigos!!!¶
Dlib detecção de rosto¶
Preparativos...
In [10]:
detector = dlib.get_frontal_face_detector()
predictor = dlib.shape_predictor('helpers/'
'shape_predictor_68_face_landmarks.dat')
O poder do Dlib¶
In [11]:
image_01 = cv2.imread('persons/bolsonaro.png')
image_01 = cv2.cvtColor(image_01, cv2.COLOR_BGR2RGB)
plt.figure(figsize=(10, 10))
plt.imshow(image_01)
plt.show()
Capturar a região do rosto¶
In [12]:
face = face_detection(image_01, detector)
plt.figure(figsize=(10, 10))
plt.imshow(face)
plt.show()
Será que continua capturando o rosto se a imagem estiver alterada?¶
In [13]:
kernel = np.ones((7, 7),np.float32)/25
result_2d_filter = cv2.filter2D(image_01,-1,kernel)
plt.figure(figsize=(10, 10))
plt.imshow(result_2d_filter)
plt.show()
In [14]:
face = face_detection(result_2d_filter, detector)
plt.figure(figsize=(10, 10))
plt.imshow(face)
plt.show()
In [15]:
rows, cols, _ = image_01.shape
M = cv2.getRotationMatrix2D((cols/2,rows/2), 45, 1)
rotate_image_01 = cv2.warpAffine(image_01,M,(cols,rows))
plt.figure(figsize=(10, 10))
plt.imshow(rotate_image_01)
plt.show()
In [16]:
face = face_detection(rotate_image_01, detector)
plt.figure(figsize=(10, 10))
plt.imshow(face)
plt.show()
E se tiver mais de um rosto?¶
In [17]:
image_02 = cv2.imread('persons/bolsonaro_02.jpg')
image_02 = cv2.cvtColor(image_02, cv2.COLOR_BGR2RGB)
plt.figure(figsize=(10, 10))
plt.imshow(image_02)
plt.show()
Será que ele detecta mesmo todos os rostos?¶
In [18]:
faces = face_detection(image_02, detector)
plt.figure(figsize=(10, 10))
plt.imshow(faces)
plt.show()
Pontos de interesse do rosto¶
Vamos ver esses pontos nas imagens¶
In [19]:
land = landmarks_detection(image_02, detector, predictor)
plt.figure(figsize=(15, 15))
plt.imshow(land)
plt.show()
In [20]:
land = landmarks_draw(image_02, detector, predictor)
plt.figure(figsize=(10, 10))
plt.imshow(land)
plt.show()
Regiões de interesse do rosto¶
In [21]:
land = facial_landmarks(image_01, predictor, detector)
plt.figure(figsize=(10, 10))
plt.imshow(land)
plt.show()
Vamos brincar um pouco com essas regiões de interesse... ou tentar.¶
In [22]:
overlay = image_01.copy()
pts = get_coordinates(image_01, 'mouth', detector, predictor)[0]
for l in range(1, len(pts)):
ptA = tuple(pts[l - 1])
ptB = tuple(pts[l])
cv2.line(overlay, ptA, ptB, (0, 255, 0), 2)
In [23]:
plt.figure(figsize=(10, 10))
plt.imshow(overlay)
plt.show()
Ops... Vamos tentar de novo...¶
In [24]:
overlay = image_01.copy()
pts = get_coordinates(image_01, 'mouth', detector, predictor)[0]
In [25]:
ptA = tuple(pts[0]) # Canto esquerdo da boca
ptB = tuple(pts[6]) # Canto direito da boca
ptC = tuple(pts[3]) # Canto superior da boca
ptD = tuple(pts[9]) # Canto inferior da boca
h = abs(ptC[1] - ptD[1])
w = abs(ptA[0] - ptB[0])
# img[y:y+h, x:x+w]
new_overlay = overlay[ptA[1]: ptA[1] + h, ptA[0]: ptA[0] + w]
In [26]:
plt.figure(figsize=(10, 10))
plt.imshow(new_overlay)
plt.show()
In [27]:
offset = 20
# img[y:y+h, x:x+w]
new_overlay = overlay[ptA[1] - offset: ptA[1] + h,
ptA[0] - offset: ptA[0] + w + offset]
plt.figure(figsize=(10, 10))
plt.imshow(new_overlay)
plt.show()
Que tal um pouco de barba?¶
In [28]:
sticker = cv2.imread('stickers/barba_02.png', -1)
watermark = cv2.resize(sticker, None, fx=0.4, fy=0.4)
plt.figure(figsize=(10, 10))
plt.imshow(watermark)
plt.show()
In [29]:
frame = cv2.cvtColor(image_01, cv2.COLOR_RGB2BGRA)
frame_h, frame_w, frame_c = frame.shape
In [30]:
overlay = frame.copy()
watermark_h, watermark_w, watermark_c = watermark.shape
In [31]:
for i in range(0, watermark_h):
for j in range(0, watermark_w):
if watermark[i, j][3] != 0:
w_offset = ptB[0] - (w*2) # ptB é o canto direito da boca
h_offset = ptD[1] - (int(ptD[1]/5) + offset) # ptD é o canto inferior da boca
overlay[h_offset + i, w_offset + j] = watermark[i, j]
overlay = cv2.cvtColor(overlay, cv2.COLOR_BGRA2RGB)
In [32]:
plt.figure(figsize=(10, 10))
plt.imshow(overlay)
plt.show()
Ou um óculos?¶
In [33]:
stickers = cv2.imread('stickers/oculos_01.png', -1)
watermark = cv2.resize(stickers, None, fx=0.6, fy=0.6)
plt.figure(figsize=(10, 10))
plt.imshow(cv2.cvtColor(watermark, cv2.COLOR_BGRA2RGB))
plt.show()
Precisamos saber onde está os olhos...¶
In [34]:
overlay = image_01.copy()
pts = get_coordinates(image_01, 'right_eye', detector, predictor)[0]
ptA = tuple(pts[0]) # Canto esquerdo do olho
ptB = tuple(pts[3]) # Canto direito do olho
ptC = tuple(pts[1]) # Canto superior do olho
ptD = tuple(pts[5]) # Canto inferior do olho
h = abs(ptC[1] - ptD[1])
w = abs(ptA[0] - ptB[0])
offset = 30
# img[y:y+h, x:x+w]
overlay = overlay[ptA[1] - offset: ptA[1] + h,
ptA[0] - offset: ptA[0] + w + offset]
In [35]:
plt.figure(figsize=(10, 10))
plt.imshow(overlay)
plt.show()
E agora vamos ajudar ele a enxergar melhor :)¶
In [36]:
frame = cv2.cvtColor(image_01, cv2.COLOR_RGB2BGRA)
frame_h, frame_w, frame_c = frame.shape
# overlay with 4 channels BGR and Alpha
overlay = frame.copy()
watermark_h, watermark_w, watermark_c = watermark.shape
# replace overlay pixels with watermark pixel values
for i in range(0, watermark_h):
for j in range(0, watermark_w):
if watermark[i,j][3] != 0:
w_offset = ptB[0] - (w*3)
h_offset = ptD[1] - (int(ptD[1]/2) + offset)
overlay[h_offset + i, w_offset + j] = watermark[i, j]
overlay = cv2.cvtColor(overlay, cv2.COLOR_BGRA2RGB)
In [37]:
plt.figure(figsize=(10, 10))
plt.imshow(overlay)
plt.show()
Vamos testar com outra imagem¶
In [38]:
image_03 = cv2.imread('persons/barack_obama.jpg')
image_03 = cv2.cvtColor(image_03, cv2.COLOR_BGR2RGB)
plt.figure(figsize=(10, 10))
plt.imshow(image_03)
plt.show()
Vamos usar aquele mesmo óculos¶
In [39]:
stickers = cv2.imread('stickers/oculos_01.png', -1)
watermark = cv2.resize(stickers, None, fx=1, fy=1)
plt.figure(figsize=(10, 10))
plt.imshow(cv2.cvtColor(watermark, cv2.COLOR_BGRA2RGB))
plt.show()
E vamos capturar a região dos olhos de novo...¶
In [40]:
overlay = image_03.copy()
pts = get_coordinates(image_03, 'right_eye', detector, predictor)[0]
ptA = tuple(pts[0]) # Canto esquerdo do olho
ptB = tuple(pts[3]) # Canto direito do olho
ptC = tuple(pts[1]) # Canto superior do olho
ptD = tuple(pts[5]) # Canto inferior do olho
h = abs(ptC[1] - ptD[1])
w = abs(ptA[0] - ptB[0])
offset = 30
# img[y:y+h, x:x+w]
overlay = overlay[ptA[1] - offset: ptA[1] + h,
ptA[0] - offset: ptA[0] + w + offset]
In [41]:
plt.figure(figsize=(10, 10))
plt.imshow(overlay)
plt.show()
Agora a melhor parte o/¶
In [42]:
frame = cv2.cvtColor(image_03, cv2.COLOR_RGB2BGRA)
frame_h, frame_w, frame_c = frame.shape
# overlay with 4 channels BGR and Alpha
overlay = frame.copy()
watermark_h, watermark_w, watermark_c = watermark.shape
# replace overlay pixels with watermark pixel values
for i in range(0, watermark_h):
for j in range(0, watermark_w):
if watermark[i,j][3] != 0:
w_offset = ptB[0] - (w*2) - offset
h_offset = ptD[1] - (int(ptD[1]/2) + offset)
overlay[h_offset + i, w_offset + j] = watermark[i, j]
overlay = cv2.cvtColor(overlay, cv2.COLOR_BGRA2RGB)
In [43]:
plt.figure(figsize=(10, 10))
plt.imshow(overlay)
plt.show()
